Context
Graphisoft · Creative Day → GRID → shipped add-on
Role
UX & UI designer · team of 3–4 designers
Duration
May–Oct 2023 · shipped Nov 2023
Scope
Ideation · User testing · UI · Dev handoff
Generative AI in Archicad, before it was easy
AI Visualizer turns a simple mass model in Archicad into images of a building, for early visuals, quick ideas and inspiration without detailed modeling. The idea started at a design team Creative Day in May 2023 and became a working demo days later at GRID, Graphisoft’s company-wide hackathon, where it won the Business category. Another designer and I designed the interface; the designer who had the idea built the AI pipeline. Six months later, it shipped as a free experimental add-on for Archicad 27.
Background
In early 2023, the generative AI hype had just begun, and we wanted to bring it into Archicad early. It wasn't easy to work with. To get Stable Diffusion running on my own computer, I typed commands into Terminal, then installed AUTOMATIC1111, a free, community-built web interface that was the most common way to use Stable Diffusion at the time. It made the model usable, but it was built for AI enthusiasts: a single screen full of technical settings. A single image took minutes, sometimes much longer, and the results were hit or miss.
Our idea was to hide all of that and build it into Archicad itself, where architects already work. No separate app, no terminal: an architect places a few boxes in the model, and the AI turns that view into an image of a building.
AUTOMATIC1111 in 2023: the kind of interface we wanted to hide.
Eight ideas, one product
Creative Day is an informal, optional day organized by our design team, where designers form their own teams. On 12 May 2023, I was one of four designers testing what AI rendering could do in Archicad: renders from sketches, from site photos and from mass models, facade variations, interiors. For GRID we picked one: inspiration from a mass model. It was the simplest to build and the most useful. Other ideas, like enhancing an existing render, made the AI hallucinate. Turning a few simple boxes into something rich and inspiring was both achievable and worth having. GRID is a four-day hackathon followed by a demo day, and by the end we had a Figma prototype laid over Archicad and a working Stable Diffusion demo. The project won the Business category.
Eight ideas tested in a day. One became the product.
Our first concept tried to do everything: interiors, styles, weather, season, time of day. For the first release we deliberately made it narrow, and the extras stayed on the list. It was meant for inspiration, two or three clicks to a result, and not for client deliverables, art styles or fine-tuning. What was left were three controls: prompt, shape fidelity and image size.
The most important change was a rename. The concept had a creativity slider, which describes what the AI does. We turned it into shape fidelity, which asks what the architect cares about: how closely should the result follow my model?
The prompt got a safety net. Whatever the user types goes into a hidden prompt that already asks for a building in photo quality, so even an empty field returns a building. It also keeps the results on topic, whatever people try to type.
GRID concept, May 2023 → shipped dialog, November 2023.
The help was on screen. Nobody saw it.
Good results needed things architects don't usually model, such as ground under the building and a tree for scale. The prototype explained this with an illustration, an example prompt and hover tips.
The other designer and I tested it with four architects. They liked the results ("you can start from this"), but most of the help went unseen:
Three of the four never found the hover tips.
Only one followed the illustration. When another tester added ground in her third try, she said: "This is cooler than the last one!"
Shape fidelity confused three of them. One set it to 0%, hoping for a more building-like result.
The biggest surprise: architects steered with the model, not the prompt. When a result missed, they changed the geometry instead of rewriting the text. That made shape fidelity the most important control in the dialog.
In the end, the best fix wasn't better onboarding. The development team changed how the model view was passed to the AI, so the AI no longer needed ground or trees to get it right. The tutorial illustration became unnecessary and never shipped: the step that people kept missing simply disappeared.
The tutorial we tested, and later didn't need.
Speaking the architect's language
A mid-summer version still used the model's own terms: Steps, CFG Scale, Model Scanner. In the final one they became iterations and prompt strength, or disappeared behind good defaults. For the developers, every control came with the exact model value behind it.
The engine needed the same care. It ran locally and could fail to start, and sometimes several people shared one. So the dialog says what went wrong and what to do next, and it warns before anyone stops a shared engine.
How it landed
AI Visualizer launched on 16 November 2023 as a free experiment. Graphisoft's announcement summed it up: "you don't need to learn prompt engineering." A press release followed, and architecture tech media such as AEC Magazine, Architosh and Engineering.com picked it up.
The community reaction was mixed, and honest:
Setup was the biggest hurdle. A 16–20 GB engine, a full system drive, failed installs on Mac. Many people never got to their first image. This was the cost of running AI locally in 2023.
Some saw no point in it. A few architects called it a "3D mood board" and complained that results drifted too far from their model.
Professionals put it to the test. An architecture firm tried it on a real project and weighed the comfort of having it inside Archicad against more advanced standalone tools.
Those who got it running used it as intended. A teacher reported that weak student models turned into "excellent results", and one user suggested renaming it "AI Inspiration", which was exactly the goal. Another tried it with no model and no prompt, and it "dreams up a building anyway": the hidden prompt doing its job.
Where it went
The experiment became a product line. In Archicad 28 AI Visualizer moved to the cloud, with no installation and included in paid plans, and our local version was retired at the end of 2024. Parts of our dialog carried over to the reworked cloud version: Lock result, Iterations (still defaulting to 35), Prompt strength and the 800×600 default size. It has had a new version with every release since, up to AI Visualizer 3.0 in Archicad 30. Today it offers a choice of current AI models, textures, editing, upscaling and a link to Archicad's AI Assistant.
Its headline feature, Image from composition, is still the idea we started with: the model is the brief, and the prompt adds the style.
What I took from it
Users talk to the AI in their own language. We designed around the prompt, but architects steered by modeling.
On screen isn't the same as seen. The advice was right; almost nobody read it. The real fix was removing the need for it.
It's still a live feature, and it keeps evolving. I worked on the experimental first version, as an optional project next to my main work. Once it became a product, the team responsible for that area took it further, the way every piece of software grows.
AI was the material, not the tool. In 2023 we designed for AI, not with it.